Project status: binary.
I had a blast designing the proposal. This can't possibly be work.
The beginning was strange. The solution seemed so obvious that someone had to have thought of it already. A deterministic referee orchestrating generative AIs in real time. And blind to the state of the object they process.
The system isn't perfect. Even though it conceptually captures the common behavioural errors of generative AI, a residual error rate remains. Realising that stopped me for a week. One-Of-Us detects correlated error at runtime; proactively predicting every case before production is an open problem. I can know where they have historically failed together, pull them from the component pool, and test them before releasing them back to production. And raise the temperature of the adversarial AI while fixing the seed. But I cannot guarantee 100% error detection — that's the truth. The rest of the auditor LLMs, the boolean modules and a good (I think) quality management system mitigate this error. But they don't eliminate it. Although, if the system feeds back on reality through human intervention, it tends toward perfection. As perfect as the reality model of the human in charge.
After validating the underlying mathematical design, I explored different theoretical exploitation routes and investigated deployment trajectories for the architecture: public and private funding, promotion on technical platforms, technology transfer, cold outreach, meaningful intervention in specialised networks, etc. I have no social media. Quirks of an old woman with cats.
That I alone have reached significant frontiers of technical depth on the behaviour of Artificial Intelligences is odd. That it sounds unhinged makes me doubt. I can neither deny nor absolutely affirm that it is. After all, this is a system built on the fact that AIs fail, validated (and knocked down along the way) by probabilistic AIs and by myself. That's the catch. I'm disconnected from the market. I live in a precise universe. So I have no contacts. I expected my work to speak for itself. But the initial difficulty of communicating its content, the methodical doubt about my own sycophancy bias and the absence of prior credibility, added together, tip the scales. It weighs on me too.
I have questioned the system's empirical isomorphism. And only a rigorous, objective PoC would show whether the increase in data integrity, the structuring of human oversight and the reduction of the error rate would justify the implementation and the added computational cost. It is hard to convey a complex problem and a proportionate solution, however didactic you try to be. So I've decided to make videos unpacking the problems I've detected in my years of exchanges with probabilistic AI. It felt like the most organic and honest move on my part. And it's what I most feel like doing after months hyperfocused on the architecture.
Everything is possible.